Automated performance debugging of production applications

Abstract

Performance anomalies in production applications can be analyzed to determine the dynamic behavior over time of hosting processes on the same or different computers. Problematic call sites (call sites that are performance bottlenecks or that are causing hangs) can be identified. Instead of relying on static code analysis and development phase load testing to identify a performance bottleneck or application hang, a lightweight sampling strategy collects predicates representing key performance data in production scenarios. Performance predicates provide information about the subject (e.g., what the performance issue is, what caused the performance issue, etc.). The data can be fed into a model based on a decision tree to identify critical threads running the problematic call sites. The results along with the key performance data can be used to build a call graph prefix binary tree for analyzing call stack patterns. Data collection, analysis and visualizations of result can be performed.